WMT20 Markables

Built with TypeScript/Webpack frontend and Python/Flask backend and processing scripts. Read the paper here and please cite it as:

@inproceedings{zouhar2020wmt20,
  title={WMT20 Document-Level Markable Error Exploration},
  author={Zouhar, Vil{\'e}m and Vojt{\v{e}}chov{\'a}, Tereza and Bojar, Ond{\v{r}}ej},
  booktitle={Proceedings of the Fifth Conference on Machine Translation},
  pages={371--380},
  year={2020},
  url={https://aclanthology.org/2020.wmt-1.41}
}

This dataset contains human annotated translation quality of terminologies/markables:

• processing/doc_data/doc_data.tar.xz: documents (sources and references and machine translations)
• collected_data/collected_data.tar.xz: collected annotations

The easiest way of accessing the data is via HuggingFace:

from datasets import load_dataset
dataset = load_dataset("zouharvi/wmt20-markables")

Development

• Interface: After git clone and npm install run npm run build, which outputs the whole frontend to interface/dist. Furthermore you can run npm run dev, which starts a local http server and serves the current project and recompiles & reloads on any code change.
• Server: To run the server simply execute the run.sh. Make sure that the url and the port matches the one specified in the interface.

Experiment preparation

To run the experiment, two files must be supplied: content.json and queue_user.json. These files are produced using provided scripts from an experiment.yaml file and a directory of documents. The json files need to be put in the backend/logs directory so that the server can use them. The output of this experiment is a file backend/logs/rating_user.json where the user ratings are stored.

The experiment.yaml file has four top-level components: users, docs, mts and markables. An example experiment.yaml file is provided in the processing/ directory.

• users is a list of user ids
• mts is a list of machine translation submission models
• docs is a list of documents to be evaluated
• markables is a list of keys, which then map to a list of corresponding forms in the source language

markables:
  uni:
    - university
    - univerzita
    - univerzit

  literature:
    - literature
    - literatura
    - literatury
    - literatuře
...

Assuming the relevant experiment.yaml file has been written and that the documents are stored in example_docs in the format of DOCNAME_MTNAME.xml you can run the following commands to generate experiment and copy it to the desctination.

processing/src/prepare_queue.py processing/experiment.yaml processing/example_docs/
processing/src/prepare_content.py processing/experiment.yaml processing/example_docs/
mv {content,queue_user}.json backend/logs/

Data preparation

Extraction

WMT .sgm to .txt was done using a script similar to this:

for file in ./*.sgm; do
	echo $file 
	tail -n +14973 "$file" | head -n 13 | sed -E 's/<seg id="[0-9]*">//g; s/<\/seg>//g' > tmp
	mv tmp "$file"
done

Model renaming

• PROMT_NMT - eTranslation: Since PROMT_NMT only does cs->en and eTranslation en->cs, the models were unified into one called PROMT_NMT-eTranslation so that it's easier to process. They should be separated in the evaluation.
• Newstest online-[abc]: The newstest2020 online models were stripped of all the language codes so that in both directions only three models could be used: newstest2020-online-[abc].sgm.

Data types

Document	Direction	Lines	Markable Occurences	Description	Source
kufrc	cs->en	29	73	housing lease agreement	SMLprodl
kufre	en->cs	29	70	housing lease agreement	SMLprodl
autoc	cs->en	18	11	Brno Grand Prix article, highway accident report	denik.cz.201550+blesk.cz.189684
euroe	en->cs	13	15	Greta Thunberg's U.N. speech article	rt.com.113881
broukc	cs->en	90	28	Supreme Audit Office audit report	KA_13_04
brouke	en->cs	90	18	Supreme Audit Office audit report	KA_13_04
Total	-	269	215	-	-